Reinforcement Learning Startups funded by Y Combinator (YC) 2026

Evaluating models for social behavior, safety, and performance through multi-agent simulated games. We use popular games like Catan, Risk, or Poker, and make humans come play them against talking AI opponents, but on the backend we're actually running a multi-agent environment used to evaluate the models for skill and behavior over large sample sizes. You can play in Social Arena and view our preliminary benchmarks today at https://olamlabs.ai/

The everyday person should own their intelligence. Our platform trains custom AI models and memory systems, hosts them privately, and continually improves them as they are used.

The promise of AI agents is still unmet in complex, real-world business environments. Our mission is to make it possible for agents to reliably accomplish concrete and useful tasks. That’s why we're building **Alexandria,** the world's largest library of real-world business workflow datasets.

Fabraix builds state-of-the-art AI red-teaming agents that continuously detect security vulnerabilities in customer-facing AI. Our product, Nyx, has already found vulnerabilities in agents at dozens of Fortune 500 companies. On AgentHarm, the leading benchmark for offensive AI security, Nyx achieved a 78% attack success rate, compared with 67% for GPT-5.6 Sol. Companies usually pay for penetration tests one project at a time. They select which systems to include, give a team a few weeks to find vulnerabilities, and receive a report when the project ends. Testing everything this way is very expensive. Globally, penetration tests cover only 26% of the software attack surface. AI agents can change more often than teams can test them. A new model, prompt, tool, permission, or data source can change what an agent does, even when the application code stays the same. The report describes only the version that was tested. AI is also increasing how much software companies produce and how often it changes. We built Nyx to automate this work. It does three things: 1. Nyx connects through the same interfaces customers use. It tests chat, voice, browser, and coding agents without source code or a special integration. 2. It draws on more than 10,000 jailbreaks that we have collected and classified, the largest such library we know of. 3. It uses each response to decide what to try next and can pursue a promising approach for hundreds of turns. In our tests, these adaptive attacks succeeded 20 times as often as attacks that were replayed unchanged or stopped after a fixed number of turns. Nyx often finds its first vulnerability within minutes or hours rather than days or weeks. Because the work is automated, companies can repeat the test with every change and at much lower cost. We're also the team behind ACE (Adversarial Cost to Exploit), a benchmark that measures AI security in terms of how much it costs attackers to break an AI system; providing a game-theoretic framework to understand how motivated a rational attacker would be in exploiting the system.

Praxis captures the data physical AI is bottlenecked on: egocentric video, 3D scans, and multimodal capture from inside real industrial and residential environments. Embedded within publicly listed and unicorn-scale conglomerates, we reach 60k workers across 5 continents and 150+ environment types, supplying the real-world training data frontier labs and humanoid companies can't source anywhere else.

We source high quality tasks and data to train the next generation of computer use AI models.

Brumby builds autonomous drones that herd cattle. On command, our drones fly to a paddock, position themselves around the mob, and move them where they need to go. What used to take a full day of helicopters, motorbikes, and horses now runs on a schedule. We work with some of the largest cattle ranches in the world. While the drones are herding, they're also estimating animal weights, measuring grass biomass, monitoring water levels, and flagging sick animals. We're building physical AI that lets a grazier manage thousands of head across millions of acres from their phone.

Cortex AI builds the world’s most diverse and large-scale real-world workplace robot & egocentric dataset — where the physical world becomes the next training and evaluation set for embodied AI. We power frontier labs developing robotics foundation models and general-purpose robots by providing the data they need: 1️⃣ Egocentric Data — real-workplace human video with hand/body pose, depth, and subtask labels. 2️⃣ Robot Data — trajectories collected from manipulators and humanoids in real industry settings. 3️⃣ Human-in-the-Loop Rollouts & Evals — real-world deployments with remote operators who recover robots when they fail, capturing data that feeds back into training and continuously improves models. Additionally, through the Cortex Marketplace, workplaces get paid to host data-collection and evaluation sessions, while labs access the in-the-wild data that truly matters. This draws on Lucas’s previous experience as co-founder of Carousell, a C2C marketplace that scaled to a $1B+ valuation.

We work with frontier AI labs to help train their agents to become AI research scientists

Topological is developing physics-based foundation models for CAD optimization. We help hardware teams iterate at the same speed that software teams do. Our technology is accelerating the engineering workflow with AI and scales design and optimization to identify the ideal designs for complex problems given their physical constraints with enhanced speed and performance. Our first model, UToP-v1, is a SOTA topology optimization model that understands physics, geometry, and manufacturability. It can generate the most efficient design given a problem’s physical requirements. It has <5% compliance error and is 1930x faster than current methods. We're reimagining mechanical engineering and computational design with precision spatial AI.

Monte is an applied research lab building products for continual learning and recursive self-improvement for AI agents.

Idler builds reinforcement learning environments that teach AI models to code at expert human levels. We create training environments based on real-world coding scenarios that prepare models for the complex challenges they'll face in production.

Kairos closes the last-mile reliability gap in AI deployments. We bring frontier techniques to companies in critical industries, deploying specialized agents that encode operator expertise and reliably automate their most manual workflows.

Aviro is a research partner with four frontier AI labs, F500 companies, and some of the top RL data vendors building post-training datasets. We focus on tasks involving thousands of tool steps for coding, computer use, and knowledge work.

We're creating reinforcement learning environments for training frontier models.

Freesolo works with companies to encode user-trajectory knowledge into specialized models that outperform the state of the art at much lower latency and cost.

HUD (YC W25) is developing agentic evals and RL environments for Computer Use Agents (CUAs) that browse the web for frontier AI labs. Our CUA Evals framework is the first comprehensive evaluation tool for CUAs. People don't actually know if AI agents are working reliably. To make AI agents work in the real world, we need detailed evals for a huge range of tasks. We're backed by Y Combinator, and work closely with frontier AI labs to provide agent evaluation and training infrastructure at scale.

At Agentin AI, we build Agents that move data and take actions across enterprise systems, like Salesforce, NetSuite and SAP. These agents are difficult to build because each enterprise heavily customizes their systems but we solved that by training our Agents to learn and adapt from failures, applying reinforcement learning techniques we developed.

TrainLoop makes it effortless for developers to supercharge LLM performance through reinforcement learning.

Osmosis is a post-training platform that helps companies fine-tune language models using reinforcement learning. We work with fast-growing AI companies to train task/domain-specific models that beat foundation models on performance, cost, and latency. Our platform handles compute orchestration, reward modeling, and training run observability as a CLI-based product usable by developers and agents.

Choose a coding agent harness, model, and task dataset and optimize context and prompts to get the best performance for long-horizon tasks

We work on methods to autonomously scaling evals/environments for post-training agents.

Always-on AI agents for 1-1 personalization at scale

Velos helps non-technical operations teams automate complex, manual back-office tasks with AI workers instead of overseas teams. Unlike traditional robotic process automation (RPA) platforms like UiPath, Velos automations use machine learning to reliably handle ambiguities in their tasks, eliminating the need for an army of maintenance engineers and consultants to build and maintain your automations. We're automating the repetitive work people hate to do.

Founded in 2022, Atmeto was started as a place to develop and apply machine learning to solve the world's biggest problem—climate change. Our current priority is getting the grid to run on 100% clean energy, which is currently limited by battery storage (specifically, the algorithms that control them). We're redefining these algorithms to unlock gigawatts of untapped energy storage capacity, enabling the grid to run on more clean energy from wind and solar.

rct AI is providing AI solutions to the game industry and building the true Metaverse with AI generated content. By using cutting-edge technologies, especially deep learning and reinforcement learning, rct AI creates a truly dynamic and intelligent user experience both on the consumers’ side and production’s side. The founding team ever built a company, Raventech together and helped make it acquired by Baidu (NASDAQ:BIDU) in 2017.

Sepal is a data research company on a mission to advance human knowledge and capabilities through safe AI. We partner with the world’s leading AI labs and enterprises to help their models get better at the tasks people actually want them to do. We’ve built a Cloud-Native Agent Dataset Factory which turns the process of generating evaluation and training data from manual, inconsistent, and labor-intensive into something automated, standardized, and scalable. Sepal AI was founded in 2024 by engineers and operators from Vercel and Turing. We went through Y Combinator, raised several million dollars from leading investors, and already count multiple Fortune 500s and top AI research labs as paying customers.

Resonance hyper-personalizes MarTech campaign content and automatically refreshes and stores high performing content for re-use.